Method and system for generating time series of hybrid diffusion model

Abstract

The invention belongs to the technical field of time prediction, and particularly relates to a time sequence generation method and system for a mixed diffusion model, and the method comprises the steps: building a dual-stage diffusion framework, dividing a diffusion process into an early stage and a late stage, and enabling each stage to correspond to a diffuser model; the dependency relationship and the dynamic change of the time sequence data are captured on the basis of a model of a Transform architecture; a diversified loss function design is utilized to ensure that the generated time sequence sample has time consistency and can maintain distribution similarity with original data; the diffuser models of the two stages are trained at the same time through a combined training mechanism; according to the method, adaptive optimization in a data scarcity scene is realized by using an adaptive optimization strategy, through a two-stage diffusion framework and a joint training mechanism, global and local features are learned under different noise levels, and diversity of generated samples under complex time sequence data is ensured.

Images